300B Tokens = 10,000 Years of Nonstop Speech: Fleuret Puts Training Scale in Perspective

francoisfleuret · x · 2026-09-09

Former Meta research scientist Francois Fleuret quantifies LLM training corpora: at 8,000 words spoken per hour, 300B tokens (230B words) equals about 10,000 years of speaking 8 hours a day — a vivid illustration of how much data modern models consume.

Original post →

More from Research

Research channel →